USPTO Automated Search Pilot Ends April 20: What It Means for Insurance AI Patents

The USPTO's Automated Search Pilot closes its intake window on April 20, 2026. It launched on October 20, 2025, targeting roughly 1,600 applications, about 200 per technology center across eight centers, each participating examiner receiving an Automated Search Results Notice ranking the ten most relevant prior art documents to the claims as filed.

For insurance AI filers the pilot is less an AI-in-government story than a controlled read on how examiners handle pricing, fraud and underwriting claim language under the revised Section 101 framework, because it ran through exactly the window in which that framework reached the examination cycle.

Key Takeaways

• The pilot targets 1,600 applications at roughly 200 per technology center, with participation by examiner selection rather than applicant opt-in, so filers learn they are in it when an ASRN appears in the file wrapper.
• Roughly 60 to 90 pilot applications across TC 3600 and TC 2100 are estimated to carry insurance-relevant claim language, a thin slice of each center's docket but a thick slice of insurance AI.
• The ASRN is not a search of record: the examiner need not adopt it, cite it, or treat it as a substitute for their own search.
• Broad machine learning claims pull back generic computer science prior art that is hard to distinguish; claims carrying insurance vocabulary pull back narrower actuarial prior art that can be attacked on the merits.
• The USPTO has not committed to publishing the share of ASRN-listed documents examiners actually cite, which is the statistic that determines whether the tool is advisory or is setting the prior art of record.

What the Pilot Is and Where It Touches Insurance

The pilot is a procedural experiment rather than a rulemaking. When a participating application enters examination, a USPTO-developed tool searches the internal prior art corpus and external databases, ranks the top ten documents by relevance to the claims, and delivers the list to the examiner. The examiner is not required to adopt or cite it and is explicitly told it does not replace their own search.

Four structural parameters were set in the October announcement. Enrollment caps at about 200 applications per technology center across eight centers, for roughly 1,600 in total. Participation is by examiner selection, not applicant opt-in. The pilot runs to April 20, 2026 with no announced extension. And the Office committed to evaluating consistency, time savings and prior art quality before deciding whether to expand, make permanent, or fold the tooling elsewhere.

Three of the eight centers absorb nearly all the relevant docket. TC 3600 routes most insurance business method filings, including pricing algorithms, underwriting workflows and claims adjudication systems, through Art Units 3693 and 3694. TC 2100 handles machine learning architecture claims and foundation model adaptations regardless of industry. TC 2600 takes a smaller but growing share from filers who frame insurance AI as communications or signal processing specifically to avoid business method treatment.

Roughly 60 to 90 pilot applications across TC 3600 and TC 2100 are estimated to carry insurance-relevant claim language. Against each center's annual docket that is thin. Against the population of insurance AI applications examined with an AI-ranked prior art list attached, it is the whole sample.

Broader Claims Now Pull Worse Prior Art

The ASRN changes nothing about the legal standard. What it changes is which prior art reaches the examiner first, and the emerging pattern in published file wrappers runs against the usual drafting instinct.

A claim reciting collection of data, training of a generic model and output of a prediction pulls back a top-ten list heavy on computer science references, academic papers and non-insurance patents. That is unhelpful prior art for an applicant trying to distinguish, and it sits squarely inside the framing that generic machine learning applied in a new environment is not eligible.

A claim describing a specific insurance workflow, a loss development calculation tied to a particular triangle method or a fraud scoring step tied to a named adjudication decision, pulls back narrower insurance-specific references. Filings using language such as loss ratio indication, credibility-weighted or schedule P triangle draw ASRNs clustering around CAS, NAIC and actuarial software prior art rather than generic ML publications.

Narrow prior art is the better outcome. It can be distinguished claim by claim; a list of foundational machine learning references cannot.

That inverts the drafting economics for the three insurance categories most exposed after Recentive Analytics. A pricing patent written as a method for determining an insurance premium using a machine learning model is maximally exposed, and it also draws the least tractable ASRN. The alternative anchors on the model rather than the use: a training methodology that handles low-credibility segments through a named constraint, or an architecture that reduces the variance of rate indications by a measurable amount. Both are actuarial statements before they are legal ones, and both give the examiner something the generic prior art does not read on.

Fraud detection sits in the hardest position, because the underlying task is pattern recognition on claims data, a canonical mental process, and ASRNs on broad fraud claims return long-standing generic anomaly detection patents. Agentic underwriting sits in the easiest, because multi-agent coordination, tool use and memory management are distinguishable at the structural level.

The timing is what makes the sample worth reading. The August 4, 2025 Squires memo narrowed the mental process grouping and directed holistic weighing of improvements. The Office rescinded the February 2024 AI inventorship guidance on November 28, 2025. Advance notice of MPEP revisions incorporating Ex parte Desjardins followed on December 5. Every one of those had to land in an examiner's workflow during the same six months the ASRNs were being issued.

The Pilot Improved the Prior Art Universe, Which Is Not Free

The uncomfortable finding is what the ASRNs are surfacing. Several pilot applications have received top-ten lists including patents held by State Farm, USAA, Allstate and AIG on adjacent claim categories, references a human first-pass search would not have reached.

No new patents were granted to produce that. The prior art universe simply got better documented, and freedom to operate narrowed for everyone who had been relying on the limits of conventional search. That falls hardest on the insurtech and services filers of the 2018 to 2022 wave, whose portfolios cluster in document extraction, knowledge graphs and regulatory reporting across the same two technology centers, and whose independence in drafting is not a defense against prior art that was always there.

The scale of the effect turns on a number the Office has not committed to publishing: the share of ASRN-listed documents examiners actually cite. Below roughly 20 percent the notice is advisory and the examiner's own search still governs. Above 50 percent the tool is materially shaping what becomes prior art of record, which reaches past examination into prosecution history estoppel and claim construction in any later dispute.

That is a considerable amount of downstream consequence resting on an adoption rate. The announced evaluation covers examiner consistency, time savings and prior art quality, and a productivity finding on those three is compatible with either reading. Filers can approximate the answer only from their own file wrappers, one continuation at a time, which is precisely the sort of asymmetry that favors the carriers already holding the densest portfolios.